RTX Spark and MXC: How Windows Is Becoming a Platform for AI Agents

Microsoft and Nvidia want to turn the Windows PC into a workplace for AI agents. At a Microsoft event in San Francisco on October 7, 2026, where Nvidia CEO Jensen Huang also appeared, the two companies presented two building blocks that belong together: the RTX Spark chip, designed to run large language models directly on laptops, and Microsoft Execution Containers, or MXC, which lock agents on Windows into controlled sandboxes. The laptops are available for preorder now, starting at 2,899 euros in Germany.

Key takeaways

• RTX Spark combines a Blackwell GPU with up to 6,144 cores and a Grace CPU with up to 20 cores; Nvidia cites up to 128 GB of unified memory and one petaflop of AI compute.
• The first laptops from Microsoft, Asus, Dell, HP, Lenovo, and MSI are available for preorder, with deliveries starting October 16 at the earliest.
• In Germany, the Surface Laptop Ultra starts at 2,899 euros, while the full version of the chip starts at 4,239 euros.
• MXC is generally available for Windows 11 and defines which files and network destinations an agent may access.
• Key management features for businesses, such as Intune support, are coming later.

What RTX Spark is supposed to deliver

At its core, RTX Spark is a combination chip: the GPU and CPU sit close together and share unified memory. According to Nvidia, the two are connected at 600 GB per second. For local AI, memory is what matters most, because a language model has to fit into it completely to run smoothly. With up to 128 GB, you can run models that an ordinary laptop simply does not have room for. Microsoft’s Windows chief, Pavan Davuluri, says the Surface Laptop Ultra can run local models with more than 120 billion parameters.

As examples, Microsoft cites its own coding model MAI Code 1.1 Flash, which runs locally with 137 billion parameters, 6.8 billion of them active at a time, compressed to three bits. DeepSeek V4 Flash with 284 billion parameters is also supposed to run on RTX Spark. The advantages are obvious: no per-request costs, no data in someone else’s data center, and work without an internet connection. We explained how this works in principle using Nvidia’s DGX Spark AI box, to which RTX Spark is technically closely related.

On speed, caution and comparisons are in order. Compared with a 16-inch MacBook Pro with M5 Pro, Microsoft and Nvidia claim up to 2.1 times faster first responses, 4.3 times faster image generation, and 6.2 times faster video generation, but they note themselves that the results depend on the test setup. No independent measurements exist yet.

What the laptops cost in Germany

The chip comes in two versions, and the differences are bigger than the marketing suggests. The cut-down variant has 18 CPU cores and 5,120 GPU cores and, according to heise, supports a maximum of 32 GB of memory. Only the full version with 20 and 6,144 cores allows 64 or 128 GB. For large local models, only the expensive variant is really interesting.

The prices reflect that. According to PC Games Hardware, the Surface Laptop Ultra starts at 2,899 euros with 24 GB of memory and a 512 GB SSD. The fully enabled chip starts at 4,239 euros, and 128 GB of memory costs 6,879 euros. In the US, the device starts at 2,599 dollars. According to heise, Asus, HP, Lenovo, and MSI listed their models with 32 GB of memory and a 1 TB SSD at an identical 3,300 euros at launch. October 16 is only the earliest delivery date: some retailers list certain devices for the second half of November. Compact desktops with RTX Spark are set to follow in November.

MXC: sandboxes for agents

The more far-reaching announcement comes from Microsoft. AI agents, meaning programs that read files, run commands, and open websites on their own, need broad permissions to be useful. That is exactly what makes them risky. Microsoft Execution Containers address this: developers and IT departments define which files and network destinations an agent may use, and Windows enforces those limits at runtime. Logan Iyer, who is responsible for the Windows platform at Microsoft, sums up the principle in one sentence: an agent cannot be its own security authority.

MXC offers several levels of isolation for this. The lightest, a process container, runs on Windows 11, macOS, and Linux. Stricter is the session container, which runs an agent under its own Windows account with its own desktop and clipboard. There is also a Linux container via WSL and, as an experimental option, a micro VM with hardware isolation. A learning mode logs which access attempts an agent makes and so helps with writing suitable rules.

MXC is already supported by OpenAI’s Codex, GitHub Copilot, OpenClaw, Replit, LM Studio, Nvidia’s OpenShell, and Unsloth. Anthropic’s Claude Code, Perplexity, and other vendors plan to follow, though Microsoft gives no dates. Managing permissions for agents is not a new issue: Apple also recently tightened Full Disk Access on the Mac.

What is still missing

For businesses, MXC is only half finished. Microsoft describes container management through Intune, separating agent activity from user activity in Microsoft Entra, and controlling local agents through Agent 365 only as “coming soon.” Anyone who wants to manage agents centrally on a corporate network will have to wait. Many of the announced Copilot features that use local models are also set to arrive only over the coming months, depending on device and market.

Bottom line: rules first, then compute

RTX Spark makes Windows laptops serious machines for large local models for the first time, but the price of a sensible configuration is well above 4,000 euros. For most readers, MXC is therefore the more important news: an operating system that sets limits for agents by default makes using them safer on any Windows 11 PC, without new hardware. If you are considering an RTX Spark laptop, wait for the first independent tests and pay close attention to the memory configuration. The chip only plays to its real strength at 64 GB and above.